Dual generative adversarial networks based on regression and neighbor characteristics

Weinan Jia, Ming Lu, Qing Shen, Chunzhi Tian, Xuyang ZhengPLoS 2024 · DOI · OpenAlex

Abstract

Imbalanced data is a pervasive issue in machine learning and pattern recognition, where the disparity in sample sizes across categories or target value ranges biases model performance toward the majority. While significant research has addressed imbalanced classification, the problem of imbalanced regression—where target continuous values are non-uniformly distributed—remains relatively unexplored. This paper proposes the RNGRU model, designed to capture both regression and neighbor characteristics of samples. By utilizing an adversarial framework (RNGAN), the model determines the optimal balance between these characteristics to generate synthetic data. Additionally, a new metric, the Correlation Change Rate (CCR), is introduced to evaluate the fidelity of generated samples.

Introduction: The Challenge of Imbalanced Regression

In standard machine learning contexts, "imbalanced data" typically refers to classification tasks where one class is significantly more frequent than others. However, a similar and equally challenging phenomenon occurs in regression: imbalanced regression. In these scenarios, the target continuous variables are not uniformly distributed, leading models to perform well in high-density regions but fail in sparse or "outlier" ranges.

The objective of the proposed research was to develop a generative model capable of augmenting these sparse regions. The primary challenge in synthetic data generation for regression is ensuring that the new samples are not merely noise but preserve the complex relationships—both local (neighbor similarities) and global (regression trends)—inherent in the original dataset.

Proposed Methodology

The authors introduced a dual-component framework to address data synthesis for regression tasks.

1. RNGRU (Regression and Neighbor Characteristic Learning)

The RNGRU model was designed to simultaneously learn two types of information from regression samples:

  • Regression Characteristics: The global relationship between input features and the continuous target variable.
  • Neighbor Characteristics: The local relationships and similarities between a sample and its nearest peers in the feature space.

2. RNGAN (Regression and Neighbor GAN)

Building on the RNGRU logic, the RNGAN employs a Generative Adversarial Network (GAN) framework. It uses a "confrontation" approach to automatically determine the proportion between regression and neighbor characteristics. This adversarial process aims to produce synthetic samples that are realistic enough to augment the dataset without introducing significant bias.

3. Correlation Change Rate (CCR)

To evaluate the quality of the generated data, the authors proposed the Correlation Change Rate (CCR). This index measures the similarity between the generated synthetic samples and the original samples, specifically focusing on the preservation of regression characteristics. The goal of the RNGAN is to minimize the similarity (to ensure diversity) while maintaining the structural integrity of the data.

Retraction and Academic Integrity

It is important to note that the original article, published in PLOS ONE, has been retracted. The publisher cited significant content overlap with previously published work as the reason for the removal of the technical body of the paper. This overlap compromises the novelty and academic integrity of the specific implementation described.

While the conceptual approach of combining neighbor and regression characteristics via GANs remains an area of active interest, researchers are encouraged to look toward established and verified techniques such as SMOGN (Synthetic Minority Over-Sampling Technique for Regression) or more recent, peer-reviewed GAN-based tabular data synthesis methods.

Conclusion

The proposed RNGAN framework attempted to solve the difficult problem of imbalanced regression by balancing local and global data features. However, due to the retraction of the original publication, the specific results and datasets used for validation are no longer part of the official scientific record. The core ideas—dual-feature learning and adversarial balancing—highlight the ongoing need for sophisticated data augmentation strategies in continuous-value prediction tasks.

Summary updated

Abstract

Imbalanced data is a problem in that the number of samples in different categories or target value ranges varies greatly. Data imbalance imposes excellent challenges to machine learning and pattern recognition. The performance of machine learning models leans to be partially towards the majority of samples in the imbalanced dataset, which will further affect the effect of the model. The imbalanced data problem includes an imbalanced categorical problem and an imbalanced regression problem. Many studies have been developed to address the issue of imbalanced classification data. Nevertheless, the imbalanced regression problem has not been well-researched. In order to solve the problem of unbalanced regression data, we define an RNGRU model that can simultaneously learn the regression characteristics and neighbor characteristics of regression samples. To obtain the most comprehensive sample information of regression samples, the model uses the idea of confrontation to determine the proportion between the regression characteristics and neighbor characteristics of the original samples. According to the regression characteristics of the regression samples, an index ccr (correlation change rate) is proposed to evaluate the similarity between the generated samples and the original samples. And on this basis, an RNGAN model is proposed to reduce the similarity between the generated samples and the original samples by using the idea of confrontation.
BibTeX
@inproceedings{jia2024dual,
  author = {Weinan Jia and Ming Lu and Qing Shen and Chunzhi Tian and Xuyang Zheng},
  title = {{Dual generative adversarial networks based on regression and neighbor characteristics}},
  booktitle = {PLoS},
  year = {2024},
  doi = {10.1371/journal.pone.0291656},
}
Figure